Parameters in file extraparams.

These options allow the user to refine the model in various ways, and do more involved analyses. The default values are probably fine to begin with. For Boolean options, type 1 for ``Yes'', or ``Use this option''; 0 for ``No'' or ``Don't use this option''.

FREQSCORR (double) Use a model where the allele frequencies are correlated. More specifically, rather than assuming a prior in which the allele frequencies in each population are independent draws from a uniform Dirichlet distribution, we start with a distribution which is centered around the mean allele frequencies in the sample. This model is more realistic for very closely related populations (where we expect the allele frequencies to be similar across populations), and can produce better clustering (section 4.3). The prior of $F_{ST}$ is set using FPRIORMEAN, and FPRIORSD. There may be a tendency to overestimate $K$ when FREQSCORR is turned on.

ONEFST (Boolean) Assume the same value of $F_{ST}$ for all populations. This is not recommended for most data, because in practice you probably expect different levels of divergence in each population. The important exception is if you are running the program for $K=2$, and you care about the actual value of $F_{ST}$, because in that case it is not really possible to estimate two values of $F_{ST}$ separately. When you're trying to estimate $K$, you should use the same model for all $K$ (I'd suggest ONEFST=0).

INFERALPHA (Boolean) Infer the value of the model parameter $\alpha$ from the data; otherwise $\alpha$ is fixed at the value ALPHA which is chosen by the user. This option is ignored under the NOADMIX model. (The prior for the ancestry vector $Q$ is Dirichlet with parameters $(\alpha,\alpha,...,\alpha)$. Small $\alpha$ implies that most individuals are essentially from one population or another, while $alpha>1$ implies that most individuals are admixed.)

POPALPHAS (Boolean) Infer a separate $\alpha$ for each population. Not recommended in most cases but may be useful for situations with asymmetric admixture.

RECOMBINE (Boolean) Use the linkage model. See section 4.2. RLOG10START sets the initial value of recombination rate r per unit distance. RLOG10MIN and RLOG10MAX set the minimum and maximum allowed values for log10r. RLOG10PROPSD sets the size of the proposed changes to log10r in each update. The front end makes some guesses about these, but some care on the part of the user in required to be sure that the values are sensible for the particular application.

COMPUTEPROB (Boolean) Print the log-likelihood of the data at each update, and estimate the probability of the data given $K$ and the model (see section 5). This is used in estimating $K$, and is also a useful diagnostic for whether the burnin is long enough. The main reason for turning this off would be to speed up the program ($\sim 10$-$15\%$).

INFERLAMBDA (Boolean) Infer a suitable value for $\lambda$. Not recommended for most analyses. $\lambda$ parameterizes the allele frequency prior, and for most data the default value of 1 seems to work pretty well. If the frequencies at most markers are very skewed towards low/high frequencies, a smaller value of $\lambda$ may potentially lead to better performance. It doesn't seem to work very well to estimate $\lambda$ at the same time as the other hyperparameters, $\alpha$ and $F$. POPSPECIFICLAMBDAS estimates a different $\lambda$ for each population.

NOADMIX (Boolean) Assume the model without admixture (Pritchard et al., 2000a). (Each individual is assumed to be completely from one of the $K$ populations.) In the output, instead of printing the average value of $Q$ as in the admixture case, the program prints the posterior probability that each individual is from each population. 1 = no admixture; 0 = model with admixture.

ADMBURNIN (int) (For use when RECOMBINE=1.) When using the linkage model, a short burnin with the admixture model (say 500 iterations) is strongly recommended in most circumstances. Without such a burnin, the linkage model often produces peculiar results. Set ${\rm ADMBURNIN} < {\rm BURNIN}$. We have dropped a related parameter (NOADMBURNIN) that was in Version 1.

USEPOPINFO (Boolean) Use prior population information to assist clustering. See also MIGRPRIOR and GENSBACK. Must have POPDATA=1.

GENSBACK (int) This corresponds to $G$ (Pritchard et al., 2000a). When using prior population information for individuals (USEPOPINFO=1), the program tests whether each individual has an immigrant ancestor in the last $G$ generations, where $G=0$ corresponds to the individual being an immigrant itself. In order to have decent power, $G$ should be set fairly small (2, say) unless the data are highly informative.

MIGRPRIOR (double) Must be in [0,1]. This is $\nu$ in Pritchard et al. (2000a). Sensible values might be in the range 0.001--0.1.

PFROMPOPFLAGONLY (Boolean) This option, new with version 2.0, makes it possible to update the allele frequencies, $P$, using only a prespecified subset of the individuals. To use this, include a POPFLAG column, and set POPFLAG=1 for individuals who should be used to update $P$, and POPFLAG=0 for individuals who should not be used to update $P$. This can be used both with, or without USEPOPINFO turned on. This option will be useful, for example, if you have a standard reference set of individuals from known populations, and then you want to estimate the ancestry of some unknown individuals. Using this option, the $q$ estimate for each unknown individual depends only on the reference set, and not on the other unknown individuals in the sample. This property is sometimes desirable.